In advanced device manufacturing processes, thin-film thickness control is a critical requirement that strongly influences product performance and overall yield. In practical industrial settings, however, surface roughness on thin-film substrates frequently arises, either by intentional design to improve device functionality or as an unavoidable consequence of process characteristics. Although spectroscopic reflectometry is widely employed due to its simple configuration and high throughput, its analytical accuracy fundamentally relies on the assumption of a laterally uniform and optically smooth surface. Consequently, reliable thickness measurement becomes difficult when the sample exhibits non-negligible roughness. In such cases, unwanted light scattering generated by the rough interface significantly complicates the forward spectral model, making conventional model-based analysis inaccurate or even infeasible. To overcome these limitations, we propose a deep learning-based spectral analysis algorithm that explicitly incorporates surface-roughness effects. The proposed model is trained to learn the complex interference patterns produced by rough surfaces and is able to analyze thin-film spectra that cannot be interpreted using traditional model-driven approaches. Experimental evaluation shows that our method increases the number of usable spectra by approximately a factor of five compared with conventional techniques, demonstrating its effectiveness for thin-film metrology on rough surfaces.
